ARIS: an autonomous research pipeline with a second model as referee
ARIS, short for Auto-Research-In-Sleep, is built for long unattended sessions: one agent surveys papers, proposes ideas, runs experiments and drafts the paper while an independent model reviews it. Codex over MCP is the default reviewer, and the authors document other model pairings. The method lives in plain Markdown skills, so it travels between agents.
Claude Code, Codex, Cursor, Trae, Antigravity, GitHub Copilot CLI, OpenClaw
What it does
idea-discovery goes from a broad direction to vetted ideas: literature survey, idea generation, novelty check and critical review.
auto-review-loop runs rounds of external review and fixes until the reviewer signs off or the round limit is reached.
experiment-bridge reads EXPERIMENT_PLAN.md, writes the experiment code, deploys it to a GPU and collects first results.
citation-audit verifies each bibliography entry exists, is attributed correctly and supports the claim where it is cited.
Extra tracks cover patents (claims-drafting, prior-art-search), grant proposals and Lean proofs.
When to use it
You want an experiment, review and fix loop to run overnight and a summary waiting in the morning.
Before submission you need to catch hallucinated authors, wrong years and citations used out of context.
You have a field but no concrete idea yet and need a literature scan plus ranked, novelty-checked ideas.
Ask your agent
Run idea-discovery on efficient fine-tuning of small vision-language models for edge devices
Audit paper/main.bib with citation-audit and list every entry that looks wrong
Run auto-review-loop on the draft in paper/, three rounds at most
Questions
How do I install ARIS?
Run npx skills add wanshuiyin/Auto-claude-code-research-in-sleep. The authors also ship a standalone ARIS-Code CLI and plugins for Claude Code and Codex CLI; the reviewer needs the codex MCP server registered once.
Does ARIS work in Cursor or Codex?
Yes. The README lists Claude Code, Codex CLI, Cursor, Trae, Antigravity, GitHub Copilot CLI and OpenClaw, and the repo keeps a separate skills-codex folder.
Do I need a Claude or OpenAI API key for ARIS?
Not necessarily. The authors say alternative model pairings, ModelScope-hosted ones included, run without a Claude or OpenAI API.
